{ "paper_id": "2021", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T03:35:48.718399Z" }, "title": "Extending Neural Keyword Extraction with TF-IDF tagset matching", "authors": [ { "first": "Boshko", "middle": [], "last": "Koloski", "suffix": "", "affiliation": {}, "email": "boshko.koloski@ijs.si" }, { "first": "Senja", "middle": [], "last": "Pollak", "suffix": "", "affiliation": {}, "email": "senja.pollak@ijs.si" }, { "first": "Matej", "middle": [], "last": "Martinc", "suffix": "", "affiliation": {}, "email": "matej.martinc@ijs.si" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Keyword extraction is the task of identifying words (or multi-word expressions) that best describe a given document and serve in news portals to link articles of similar topics. In this work, we develop and evaluate our methods on four novel data sets covering lessrepresented, morphologically-rich languages in European news media industry (Croatian, Estonian, Latvian, and Russian). First, we perform evaluation of two supervised neural transformer-based methods, Transformerbased Neural Tagger for Keyword Identification (TNT-KID) and Bidirectional Encoder Representations from Transformers (BERT) with an additional Bidirectional Long Short-Term Memory Conditional Random Fields (BiLSTM CRF) classification head, and compare them to a baseline Term Frequency-Inverse Document Frequency (TF-IDF) based unsupervised approach. Next, we show that by combining the keywords retrieved by both neural transformer-based methods and extending the final set of keywords with an unsupervised TF-IDF based technique, we can drastically improve the recall of the system, making it appropriate for usage as a recommendation system in the media house environment.", "pdf_parse": { "paper_id": "2021", "_pdf_hash": "", "abstract": [ { "text": "Keyword extraction is the task of identifying words (or multi-word expressions) that best describe a given document and serve in news portals to link articles of similar topics. In this work, we develop and evaluate our methods on four novel data sets covering lessrepresented, morphologically-rich languages in European news media industry (Croatian, Estonian, Latvian, and Russian). First, we perform evaluation of two supervised neural transformer-based methods, Transformerbased Neural Tagger for Keyword Identification (TNT-KID) and Bidirectional Encoder Representations from Transformers (BERT) with an additional Bidirectional Long Short-Term Memory Conditional Random Fields (BiLSTM CRF) classification head, and compare them to a baseline Term Frequency-Inverse Document Frequency (TF-IDF) based unsupervised approach. Next, we show that by combining the keywords retrieved by both neural transformer-based methods and extending the final set of keywords with an unsupervised TF-IDF based technique, we can drastically improve the recall of the system, making it appropriate for usage as a recommendation system in the media house environment.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Keywords are words (or multi-word expressions) that best describe the subject of a document, effectively summarise it and can also be used in several document categorization tasks. In online news portals, keywords help with efficient retrieval of articles when needed. Similar keywords characterise articles of similar topics, which can help editors to link related articles, journalists to find similar articles and readers to retrieve articles of interest when browsing the portals. For journalists manually assigning tags (keywords) to articles represents a demanding task, and high-quality automated keyword extraction shows to be one of components in news digitalization process that many media houses seek for.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The task of keyword extraction can generally be tackled in an unsupervised way, i.e., by relying on frequency based statistical measures (Campos et al., 2020) or graph statistics (\u0160krlj et al., 2019) , or with a supervised keyword extraction tool, which requires a training set of sufficient size and from appropriate domain. While supervised methods tend to work better due to their ability to adapt to a specifics of the syntax, semantics, content, genre and keyword assignment regime of a specific text (Martinc et al., 2020a) , their training for some less resource languages is problematic due to scarcity of large manually annotated resources. For this reason, studies about supervised keyword extraction conducted on less resourced languages are still very rare. To overcome this research gap, in this paper we focus on supervised keyword extraction on three less resourced languages, Croatian, Latvian, and Estonian, and one fairly well resourced language (Russian) and conduct experiments on data sets of media partners in the EMBEDDIA project 1 . The code for the experiments is made available on GitHub under the MIT license 2 .", "cite_spans": [ { "start": 137, "end": 158, "text": "(Campos et al., 2020)", "ref_id": "BIBREF1" }, { "start": 179, "end": 199, "text": "(\u0160krlj et al., 2019)", "ref_id": null }, { "start": 506, "end": 529, "text": "(Martinc et al., 2020a)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In media house environments, automatic keyword extraction systems are expected to return a diverse list of keyword candidates (of constant length), which is then inspected by a journalist who manually selects appropriate candidates. While the state-of-the-art supervised approaches in most cases offer good enough precision for this type of usage as a recommendation system, the recall of these systems is nevertheless problematic. Supervised systems learn how many keywords should be returned for each news article on the gold standard train set, which generally contains only a small amount of manually approved candidates for each news article. For example, among the datasets used in our experiments (see Section 3), the Russian train set contains the most (on average 4.44) present keywords (i.e., keywords which appear in the text of the article and can be used for training of the supervised models) per article, while the Croatian test set contains only 1.19 keywords per article. This means that for Croatian, the model will learn to return around 1.19 keywords for each article, which is not enough.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "To solve this problem we show that we can improve the recall of the existing supervised keyword extraction system by:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Proposing an additional TF-IDF tagset matching technique, which finds additional keyword candidates by ranking the words in the news article that have appeared in the predefined keyword set containing words from the gold standard train set. The new hybrid system first checks how many keywords were returned by the supervised approach and if the number is smaller than needed, the list is expanded by the best ranked keywords returned by the TF-IDF based extraction system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "\u2022 Combining the outputs of several state-of-theart supervised keyword extraction approaches.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The rest of this work is structured as follows: Section 2 presents the related work, while Section 3 describes the datasets on which we evaluate our method. Section 4 describes our proposed method with all corresponding steps. The experiment settings are described in Section 5 and the evaluation of the proposed methods is shown in Section 6. The conclusions and the proposed further work are presented in Section 7.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Many different approaches have been developed to tackle the problem of extracting keywords. The early approaches, such as KP-MINER (El-Beltagy and Rafea, 2009) and RAKE (Rose et al., 2010) rely on unsupervised techniques which employ frequency based metrics for extraction of keywords from text. Formally, aforementioned approaches search for the words w from vocabulary V that maximize a given metric h for a given text t:", "cite_spans": [ { "start": 122, "end": 159, "text": "KP-MINER (El-Beltagy and Rafea, 2009)", "ref_id": null }, { "start": 169, "end": 188, "text": "(Rose et al., 2010)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "kw = argmax w\u2208V h(w, t).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "In these approaches, frequency is of high relevance and it is assumed that the more frequent a given word, the more important the meaning this word carries for a given document. Most popular such metrics are the na\u00efve frequency (word count) and the term frequency-inverse document frequency (TF-IDF) (Salton and McGill, 1986) .", "cite_spans": [ { "start": 300, "end": 325, "text": "(Salton and McGill, 1986)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Most recent state-of-the-art statistical approaches, such as YAKE (Campos et al., 2020) , also employ frequency based features, but combine them with other features such as casing, position, relatedness to context and dispersion of a specific term in order to derive a final score for each keyword candidate.", "cite_spans": [ { "start": 66, "end": 87, "text": "(Campos et al., 2020)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Another line of research models this problem by exploiting concepts from graph theory. Approaches, such as TextRank (Mihalcea and Tarau, 2004) , Single Rank (Wan and Xiao, 2008) , Topi-cRank (Bougouin et al., 2013) and Topical PageRank (Sterckx et al., 2015 ) build a graph G, i.e., a mathematical construct described by a set of vertexes V and a set of edges E connecting two vertices. In one of the most recent approaches called RaKUn (\u0160krlj et al., 2019), a directed graph is constructed from text, where vertexes V and two words w i , w i+1 are linked if they appear following one another. Keywords are ranked by a shortest path-based metric from graph theory -the load centrality.", "cite_spans": [ { "start": 116, "end": 142, "text": "(Mihalcea and Tarau, 2004)", "ref_id": "BIBREF11" }, { "start": 157, "end": 177, "text": "(Wan and Xiao, 2008)", "ref_id": "BIBREF22" }, { "start": 191, "end": 214, "text": "(Bougouin et al., 2013)", "ref_id": "BIBREF0" }, { "start": 236, "end": 257, "text": "(Sterckx et al., 2015", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "The task of keyword extraction can also be tackled in a supervised way. One of the first supervised approaches was an algorithm named KEA (Witten et al., 2005) , which uses only TF-IDF and the term's position in the text as features for term identification. More recent neural approaches to keyword detection consider the problem as a sequence-tosequence generation task (Meng et al., 2017) and employ a generative model for keyword prediction with a recurrent encoder-decoder framework and an attention mechanism capable of detecting keywords in the input text sequence whilst also potentially finding keywords that do not appear in the text.", "cite_spans": [ { "start": 138, "end": 159, "text": "(Witten et al., 2005)", "ref_id": "BIBREF23" }, { "start": 371, "end": 390, "text": "(Meng et al., 2017)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Finally, the newest branch of models consider keyword extraction as a sequence labelling task and tackle keyword detection with transformers. Sahrawat et al. (2020) fed contextual embeddings generated by several transformer models (BERT (Devlin et al., 2018) , RoBERTa (Liu et al., 2019) , GPT-2 (Radford et al., 2019) , etc.) into two types of neural architectures, a bidirectional Long shortterm memory network (BiLSTM) and a BiLSTM network with an additional Conditional random fields layer (BiLSTM-CRF). Sun et al. 2020on the other hand proposed BERT-JointKPE that employs a chunking network to identify phrases and a ranking network to learn their salience in the document. By training BERT jointly on the chunking and ranking tasks the model manages to establish balance between the estimation of keyphrase quality and salience.", "cite_spans": [ { "start": 142, "end": 164, "text": "Sahrawat et al. (2020)", "ref_id": "BIBREF15" }, { "start": 237, "end": 258, "text": "(Devlin et al., 2018)", "ref_id": "BIBREF2" }, { "start": 269, "end": 287, "text": "(Liu et al., 2019)", "ref_id": "BIBREF7" }, { "start": 290, "end": 318, "text": "GPT-2 (Radford et al., 2019)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "Another state-of-the-art transformer based approach is TNT-KID (Transformer-based Neural Tagger for Keyword Identification) (Martinc et al., 2020a) , which does not rely on pretrained language models such as BERT, but rather allows the user to train their own language model on the appropriate domain. The study shows that smaller unlabelled domain specific corpora can be successfully used for unsupervised pretraining, which makes the proposed approach easily transferable to low-resource languages. It also proposes several modifications to the transformer architecture in order to adapt it for a keyword extraction task and improve performance of the model.", "cite_spans": [ { "start": 124, "end": 147, "text": "(Martinc et al., 2020a)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "We conducted experiments on datasets containing news in four languages; Latvian, Estonian, Russian, and Croatian. Latvian, Estonian and Russian datasets contain news from the Ekspress Group, specifically from Estonian Ekspress Meedia (news in Estonian and Russian) and from Latvian Delfi (news in Latvian and Russian). The dataset statistics are presented in Table 2 , and the datasets (Pollak et al., 2021) and their train/test splits 3 are publicly available. The media-houses provided news articles from 2015 up to the 2019. We divided them into training and test sets. For the Latvian, Estonian, and Russian training sets, we used the articles from 2018, while for the test set the articles from 2019 were used. For Croatian, the articles from 2019 are arranged by date and split into training and test (i.e., about 10% of the 2019 articles with the most recent date) set. In our study, we also use tagsets of keywords. Tagset corresponds either to a collection of keywords maintained by editors of a media house (see e.g. Estonian tagset), or to a tagset constructed from assigned keywords from articles available in the training set. The type of tagset and the number of unique tags for each language are listed in Table 1 ", "cite_spans": [ { "start": 386, "end": 407, "text": "(Pollak et al., 2021)", "ref_id": null } ], "ref_spans": [ { "start": 359, "end": 366, "text": "Table 2", "ref_id": "TABREF3" }, { "start": 1221, "end": 1228, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Data Description", "sec_num": "3" }, { "text": "The recent supervised neural methods are very precise, but, as was already mentioned in Section 1, in same cases they do not return a sufficient number of keywords. This is due to the fact that the methods are trained on the training data with a low number of gold standard keywords (as it can be seen from Table 2 ). To meet the media partners' needs, we designed a method that complements state-of-theart neural methods (the TNT-KID method (Martinc et al., 2020b) and the transformer-based method proposed by Sahrawat et al. (2020) , which are both described in Section 2) by a tagset matching approach, returning constant number of keywords (k=10).", "cite_spans": [ { "start": 442, "end": 465, "text": "(Martinc et al., 2020b)", "ref_id": "BIBREF9" }, { "start": 511, "end": 533, "text": "Sahrawat et al. (2020)", "ref_id": "BIBREF15" } ], "ref_spans": [ { "start": 307, "end": 314, "text": "Table 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Methodology", "sec_num": "4" }, { "text": "Both supervised neural approaches employed in this study are based on the Transformer architecture (Vaswani et al., 2017) , which was somewhat adapted for the specific task at hand. Both models are fed lowercased text consisting of the title and the body of the article. Tokenization is conducted by either using the default BERT tokenizer (when BERT is used) or by employing Sentencepiece tokenizer (Kudo and Richardson, 2018) 2018), TNT-KID requires an additional language model pretraining on the domain specific corpus.", "cite_spans": [ { "start": 99, "end": 121, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF20" }, { "start": 400, "end": 427, "text": "(Kudo and Richardson, 2018)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Transformer-based Keyword Extraction", "sec_num": "4.1" }, { "text": "In our approach, we first take the keywords returned by a neural keyword extraction method and next complement the returned keyword list by adding the missing keywords to achieve the set goal of k keywords. The added keywords are selected by taking the top-ranked candidates from the TF-IDF tagset matching extraction conducted on the preprocessed news articles and keywords.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF(tm) Tagset Matching", "sec_num": "4.2" }, { "text": "First, we concatenate the body and the title of the article. After that we lowercase the text and remove stopwords. Finally, the text is tokenized and lemmatized with the Lemmagen3 lemmatizer (Jur\u0161i\u010d et al., 2010) , which supports lemmatization for all the languages except Latvian. For Latvian we use the LatvianStemmer 4 . For the stopword removal we used the Stopwords-ISO 5 Python library which contained stopwords for all four languages. The final cleaned textual input consists of the concatenation of all of the preprocessed words from the document. We apply the same preprocessing procedure on the predetermined tagsets for each language. The preprocessing procedure is visualized in Figure 1 . 4 https://github.com/rihardsk/LatvianS temmer 5 https://github.com/stopwords-iso", "cite_spans": [ { "start": 192, "end": 213, "text": "(Jur\u0161i\u010d et al., 2010)", "ref_id": "BIBREF5" } ], "ref_spans": [ { "start": 692, "end": 700, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Preprocessing", "sec_num": "4.2.1" }, { "text": "The TF-IDF weighting scheme (Salton and McGill, 1986) assigns each word its weight w based on the frequency of the word in the document (term frequency) and the number of documents the word appears in (inverse document frequency). More specifically, TF-IDF is calculated with the following equation:", "cite_spans": [ { "start": 28, "end": 53, "text": "(Salton and McGill, 1986)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "T F \u2212 IDF i = tf i,j \u2022 log e ( |D| df i )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "The formula has two main components:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "\u2022 Term-frequency (tf) that counts the number of appearances of a word in the document (in the equation above, tf i,j denotes the number of occurrences of the word i in the document j)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "\u2022 Inverse-document-frequency (idf) ensures that words appearing in more documents are assigned lower weights (in the formula above df i is the number of documents containing word i and |D| denotes the number of documents).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "The assumption is that words with a higher TF-IDF value are more likely to be keywords.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TF-IDF Weighting Scheme", "sec_num": "4.2.2" }, { "text": "For a given neural keyword extraction method N, and for each document d, we select l best ranked keywords according to the TF-IDF(tm), which appear in the keyword tagset for each specific dataset. Here, l corresponds to km, where k = 10 and m corresponds to the number of keywords returned by a neural method.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Tagset Matching Keyword Expansion", "sec_num": "4.3" }, { "text": "Since some of the keywords in the tagsets provided by the media partners were variations of the same root word (i.e., keywords are not lemmatized), we created a mapping from a root word (i.e., a word lemma or a stem) to a list of possible variations in the keyword dataset. For example, a word 'riigieksam' ('exam') appearing in the article, could be mapped to three tags in the tagset by the Estonian media house with the same root form 'riigieksam': 'riigieksamid', 'riigieksamide' and 'riigieksam'.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Tagset Matching Keyword Expansion", "sec_num": "4.3" }, { "text": "We tested several strategies for mapping the occurrence of a word in the news article to a specific tag in the tagset. For each lemma that mapped to multiple tags, we tested returning a random tag, a tag with minimal length and a tag of maximal length. In the final version, we opted to return the tag with the minimal length, since this tag corresponded to the lemma of the word most often.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Tagset Matching Keyword Expansion", "sec_num": "4.3" }, { "text": "We conducted experiments on the datasets described in Section 3. We evaluate the following methods and combinations of methods:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 TF-IDF(tm): Here, we employ the preprocessing and TF-IDF-based weighting of keywords described in Section 4 and select the top-ranked keywords that are present in the tagset.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 TNT-KID (Martinc et al., 2020b) : For each dataset, we first pretrain the model with an autoregressive language model objective. After that, the model is fine-tuned on the same train set for the keyword extraction task. Sequence length was set to 256, embedding size to 512 and batch size to 8, and we employ the same preprocessing as in the original study (Martinc et al., 2020b ).", "cite_spans": [ { "start": 10, "end": 33, "text": "(Martinc et al., 2020b)", "ref_id": "BIBREF9" }, { "start": 359, "end": 381, "text": "(Martinc et al., 2020b", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 BERT + BiLSTM-CRF (Sahrawat et al., 2020) : We employ an uncased multilingual BERT 6 model with an embedding size of 768 and 12 attention heads, with an additional BiLSTM-CRF token classification head, same as in Sahrawat et al. (2020) .", "cite_spans": [ { "start": 20, "end": 43, "text": "(Sahrawat et al., 2020)", "ref_id": "BIBREF15" }, { "start": 215, "end": 237, "text": "Sahrawat et al. (2020)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 TNT-KID & BERT + BiLSTM-CRF: We extracted keywords with both of the methods and complemented the TNT-KID extracted keywords with the BERT + BiLSTM-CRF extracted keywords in order to retrieve more keywords. Duplicates (i.e., keywords extracted by both methods) are removed.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 TNT-KID & TF-IDF: If the keyword set extracted by TNT-KID contains less than 10 keywords, it is expanded with keywords retrieved with the proposed TF-IDF(tm) approach, i.e., best ranked keywords according to TF-IDF, which do not appear in the keyword set extracted by TNT-KID.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 BERT + BiLSTM-CRF & TF-IDF: If the keyword set extracted by BERT + BiLSTM-CRF contains less than 10 keywords, it is expanded with keywords retrieved with the proposed TF-IDF(tm) approach, i.e., best ranked keywords according to TF-IDF, which do not appear in the keyword set extracted by BERT + BiLSTM-CRF.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "\u2022 TNT-KID & BERT + BiLSTM-CRF & TF-IDF: the keyword set extracted with the TNT-KID is complemented by keywords extracted with BERT + BiLSTM-CRF (duplicates are removed). If after the expansion the keyword set still contains less than 10 keywords, it is expanded again, this time with keywords retrieved by the TF-IDF(tm) approach.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "For TNT-KID, which is the only model that requires language model pretraining, language models were trained on train sets in Table 2 for up to ten epochs. Next, TNT-KID and BERT + BiLSTM-CRF were fine-tuned on the training datasets, which were randomly split into 80 percent of documents used for training and 20 percent of documents used for validation. The documents containing more than 256 tokens are truncated, while the documents containing less than 256 tokens are padded with a special < pad > token at the end. We fine-tuned each model for a maximum of 10 epochs and after each epoch the trained model was tested on the documents chosen for validation. The model that showed the best performance on this set of validation documents (in terms of F@10 score) was used for keyword detection on the test set.", "cite_spans": [], "ref_spans": [ { "start": 125, "end": 132, "text": "Table 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Experimental Settings", "sec_num": "5" }, { "text": "For evaluation, we employ precision, recall and F1 score. While F1@10 and recall@10 are the most relevant metrics for the media partners, we also report precision@10, precision@5, recall@5 and F1@5. Only keywords which appear in a text (present keywords) were used as a gold standard, since we only evaluate approaches for keyword tagging that are not capable of finding keywords which do not appear in the text. Lowercasing and lemmatization (stemming in the case of Latvian) are performed on both the gold standard and the Model P@5 R@5 F1@5 P@10 R@10 F1@10 Croatian TF-IDF 0. extracted keywords (keyphrases) during the evaluation. The results of the evaluation on all four languages are listed in Table 3 .", "cite_spans": [], "ref_spans": [ { "start": 700, "end": 707, "text": "Table 3", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Evaluation", "sec_num": "6" }, { "text": "Results suggest, that neural approaches, TNT-KID and BERT+BiLSTM-CRF offer comparable performance on all datasets but nevertheless achieve different results for different languages. TNT-KID outperforms BERT-BiLSTM-CRF model according to all the evaluation metrics on the Estonian and Russian news dataset. It also outperforms all other methods in terms of precision and F1 score. On the other hand, BERT+BiLSTM-CRF performs better on the Croatian dataset in terms of precision and F1-score. On Latvian TNT-KID achieves top results in terms of F1, while BERT+BiLSTM-CRF offers better precision.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation", "sec_num": "6" }, { "text": "Even though the TF-IDF tagset matching method performs poorly on its own, we can nevertheless drastically improve the recall@5 and the recall@10 of both neural systems, if we expand the keyword tag sets returned by the neural methods with the TF-IDF ranked keywords. The improvement is substantial and consistent for all datasets, but it nevertheless comes at the expanse of the lower precision and F1 score. This is not surprising, since the final expanded keyword set always returns 10 keywords, i.e., much more than the average number of present gold standard keywords in the media partner datasets (see Table 2 ), which badly affects the precision of the approach. Nevertheless, since for a journalist a manual inspection of 10 keyword candidates per article and manual selection of good candidates (e.g., by clicking on them) still requires less time than the manual selection of keywords from an article, we argue that the improvement of recall at the expanse of the precision is a good trade off, if the system is intended to be used as a recommendation system in the media house environment.", "cite_spans": [], "ref_spans": [ { "start": 607, "end": 614, "text": "Table 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Evaluation", "sec_num": "6" }, { "text": "Combining keywords returned by TNT-KID and BERT + BiLSTM-CRF also consistently improves recall, but again at the expanse of lower precision and F1 score. Overall, for all four languages, the best performing method in terms of recall is the TNT-KID & BERT + BiLSTM-CRF & TF-IDF(tm).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation", "sec_num": "6" }, { "text": "In this work, we tested two state-of-the-art neural approaches for keyword extraction, TNT-KID (Martinc et al., 2020a) and BERT BiLSTM-CRF (Sahrawat et al., 2020) , on three less resourced European languages, Estonian, Latvian, Croatian, as well as on Russian. We also proposed a tagset based keyword expansion approach, which drastically improves the recall of the method, making it more suitable for the application in the media house environment.", "cite_spans": [ { "start": 95, "end": 118, "text": "(Martinc et al., 2020a)", "ref_id": "BIBREF8" }, { "start": 139, "end": 162, "text": "(Sahrawat et al., 2020)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "Our study is one of the very few studies where supervised keyword extraction models were employed on several less resourced languages. The results suggest that these models perform well on languages other than English and could also be successfully leveraged for keyword extraction on morphologically rich languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "The focus of the study was whether we can improve the recall of the supervised models, in order to make them more useful as recommendation systems in the media house environment. Our method manages to increase the number of retrieved keywords, which drastically improves the recall for all languages. For example, by combing all neural methods and the TF-IDF based approach, we improve on the recall@10 achieved by the best performing neural model, TNT-KID, by 16.81 percentage points for Croatian, 9.70 percentage points for Estonian, 9.63 percentage points for Russian and 17.12 percentage points for Latvian. The resulting method nevertheless offers lower precision, which we will try to improve in the future work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "In the future we also plan to perform a qualitative evaluation of our methods by journalists from the media houses. Next, we plan to explore how adding background knowledge from knowledge databases -lexical (e.g. Wordnet (Fellbaum, 1998)) or factual (e.g. WikiData (Vrande\u010di\u0107 and Kr\u00f6tzsch, 2014) ) would benefit the aforementioned methods.", "cite_spans": [ { "start": 221, "end": 238, "text": "(Fellbaum, 1998))", "ref_id": "BIBREF4" }, { "start": 265, "end": 295, "text": "(Vrande\u010di\u0107 and Kr\u00f6tzsch, 2014)", "ref_id": "BIBREF21" } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "The assumption is that with the linkage of the text representation and the background knowledge we would achieve a more representative understanding of the articles and the concepts appearing in them, which would result in a more successful keyword extraction.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "In traditional machine-learning setting a common practice of combining different classifier outputs to a single output is referred to as stacking. We propose further research on this topic by testing combinations of various keyword extraction models. Finally, we also plan to further improve our unsupervised TF-IDF based keyword extraction method. One way to to do this would be to add the notion of positional encoding, since some of the keywords in the news-media domain often can be found at the beginning of the article and the TF-IDF(tm) does not take this into account while applying the weighting on the matched terms.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "7" }, { "text": "http://embeddia.eu/ 2 https://github.com/bkolosk1/Extendin g-Neural-Keyword-Extraction-with-TF-IDFtagset-matching/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://www.clarin.si/repository/xml ui/handle/11356/1403", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "More specifically, we use the 'bert-base-multilingualuncased' implementation of BERT from the Transformers library (https://github.com/huggingface/tra nsformers).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This paper is supported by European Union's Horizon 2020 research and innovation programme under grant agreement No. 825153, project EM-BEDDIA (Cross-Lingual Embeddings for Less-Represented Languages in European News Media). The third author was financed via young research ARRS grant. Finally, the authors acknowledge the financial support from the Slovenian Research Agency for research core funding for the programme Knowledge Technologies (No. P2-0103) and the project TermFrame -Terminology and Knowledge Frames across Languages (No. J6-9372).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": "8" } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Topicrank: Graph-based topic ranking for keyphrase extraction", "authors": [ { "first": "Adrien", "middle": [], "last": "Bougouin", "suffix": "" }, { "first": "Florian", "middle": [], "last": "Boudin", "suffix": "" }, { "first": "B\u00e9atrice", "middle": [], "last": "Daille", "suffix": "" } ], "year": 2013, "venue": "International joint conference on natural language processing (IJCNLP)", "volume": "", "issue": "", "pages": "543--551", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adrien Bougouin, Florian Boudin, and B\u00e9atrice Daille. 2013. Topicrank: Graph-based topic ranking for keyphrase extraction. In International joint con- ference on natural language processing (IJCNLP), pages 543-551.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Yake! keyword extraction from single documents using multiple local features", "authors": [ { "first": "Ricardo", "middle": [], "last": "Campos", "suffix": "" }, { "first": "V\u00edtor", "middle": [], "last": "Mangaravite", "suffix": "" }, { "first": "Arian", "middle": [], "last": "Pasquali", "suffix": "" }, { "first": "Al\u00edpio", "middle": [], "last": "Jorge", "suffix": "" }, { "first": "C\u00e9lia", "middle": [], "last": "Nunes", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Jatowt", "suffix": "" } ], "year": 2020, "venue": "Information Sciences", "volume": "509", "issue": "", "pages": "257--289", "other_ids": { "DOI": [ "10.1016/j.ins.2019.09.013" ] }, "num": null, "urls": [], "raw_text": "Ricardo Campos, V\u00edtor Mangaravite, Arian Pasquali, Al\u00edpio Jorge, C\u00e9lia Nunes, and Adam Jatowt. 2020. Yake! keyword extraction from single documents using multiple local features. Information Sciences, 509:257 -289.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Bert: Pre-training of deep bidirectional transformers for language understanding", "authors": [ { "first": "Jacob", "middle": [], "last": "Devlin", "suffix": "" }, { "first": "Ming-Wei", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Kenton", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1810.04805" ] }, "num": null, "urls": [], "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Kpminer: A keyphrase extraction system for english and arabic documents", "authors": [ { "first": "R", "middle": [], "last": "Samhaa", "suffix": "" }, { "first": "Ahmed", "middle": [], "last": "El-Beltagy", "suffix": "" }, { "first": "", "middle": [], "last": "Rafea", "suffix": "" } ], "year": 2009, "venue": "Inf. Syst", "volume": "34", "issue": "1", "pages": "132--144", "other_ids": { "DOI": [ "10.1016/j.is.2008.05.002" ] }, "num": null, "urls": [], "raw_text": "Samhaa R. El-Beltagy and Ahmed Rafea. 2009. Kp- miner: A keyphrase extraction system for english and arabic documents. Inf. Syst., 34(1):132-144.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "WordNet: An Electronic Lexical Database", "authors": [ { "first": "Christiane", "middle": [], "last": "Fellbaum", "suffix": "" } ], "year": 1998, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. Bradford Books.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Lemmagen: Multilingual lemmatisation with induced ripple-down rules", "authors": [ { "first": "Matja\u017e", "middle": [], "last": "Jur\u0161i\u010d", "suffix": "" }, { "first": "Igor", "middle": [], "last": "Mozeti\u010d", "suffix": "" }, { "first": "Toma\u017e", "middle": [], "last": "Erjavec", "suffix": "" }, { "first": "Nada", "middle": [], "last": "Lavra\u010d", "suffix": "" } ], "year": 2010, "venue": "Journal of Universal Computer Science", "volume": "16", "issue": "9", "pages": "1190--1214", "other_ids": {}, "num": null, "urls": [], "raw_text": "Matja\u017e Jur\u0161i\u010d, Igor Mozeti\u010d, Toma\u017e Erjavec, and Nada Lavra\u010d. 2010. Lemmagen: Multilingual lemmatisa- tion with induced ripple-down rules. Journal of Uni- versal Computer Science, 16(9):1190-1214.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing", "authors": [ { "first": "Taku", "middle": [], "last": "Kudo", "suffix": "" }, { "first": "John", "middle": [], "last": "Richardson", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1808.06226" ] }, "num": null, "urls": [], "raw_text": "Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tok- enizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Roberta: A robustly optimized bert pretraining approach", "authors": [ { "first": "Yinhan", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Myle", "middle": [], "last": "Ott", "suffix": "" }, { "first": "Naman", "middle": [], "last": "Goyal", "suffix": "" }, { "first": "Jingfei", "middle": [], "last": "Du", "suffix": "" }, { "first": "Mandar", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Danqi", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Omer", "middle": [], "last": "Levy", "suffix": "" }, { "first": "Mike", "middle": [], "last": "Lewis", "suffix": "" }, { "first": "Luke", "middle": [], "last": "Zettlemoyer", "suffix": "" }, { "first": "Veselin", "middle": [], "last": "Stoyanov", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1907.11692" ] }, "num": null, "urls": [], "raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Tnt-kid: Transformer-based neural tagger for keyword identification", "authors": [ { "first": "Matej", "middle": [], "last": "Martinc", "suffix": "" }, { "first": "Senja", "middle": [], "last": "Bla\u017e\u0161krlj", "suffix": "" }, { "first": "", "middle": [], "last": "Pollak", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2003.09166" ] }, "num": null, "urls": [], "raw_text": "Matej Martinc, Bla\u017e\u0160krlj, and Senja Pollak. 2020a. Tnt-kid: Transformer-based neural tag- ger for keyword identification. arXiv preprint arXiv:2003.09166.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Tnt-kid: Transformer-based neural tagger for keyword identification", "authors": [ { "first": "Matej", "middle": [], "last": "Martinc", "suffix": "" }, { "first": "Senja", "middle": [], "last": "Bla\u017e\u0161krlj", "suffix": "" }, { "first": "", "middle": [], "last": "Pollak", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Matej Martinc, Bla\u017e\u0160krlj, and Senja Pollak. 2020b. Tnt-kid: Transformer-based neural tagger for key- word identification.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Textrank: Bringing order into text", "authors": [ { "first": "Rada", "middle": [], "last": "Mihalcea", "suffix": "" }, { "first": "Paul", "middle": [], "last": "Tarau", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 2004 conference on empirical methods in natural language processing", "volume": "", "issue": "", "pages": "404--411", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rada Mihalcea and Paul Tarau. 2004. Textrank: Bring- ing order into text. In Proceedings of the 2004 con- ference on empirical methods in natural language processing, pages 404-411.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Hannu Toivonen, Emanuela Boros, Jose Moreno, and Antoine Doucet. 2021. EMBEDDIA tools, datasets and challenges: Resources and hackathon contributions", "authors": [ { "first": "Senja", "middle": [], "last": "Pollak", "suffix": "" }, { "first": "Marko", "middle": [], "last": "Robnik\u0161ikonja", "suffix": "" }, { "first": "Matthew", "middle": [], "last": "Purver", "suffix": "" }, { "first": "Michele", "middle": [], "last": "Boggia", "suffix": "" }, { "first": "Ravi", "middle": [], "last": "Shekhar", "suffix": "" }, { "first": "Marko", "middle": [], "last": "Pranji\u0107", "suffix": "" }, { "first": "Salla", "middle": [], "last": "Salmela", "suffix": "" }, { "first": "Ivar", "middle": [], "last": "Krustok", "suffix": "" }, { "first": "Tarmo", "middle": [], "last": "Paju", "suffix": "" }, { "first": "Carl-Gustav", "middle": [], "last": "Linden", "suffix": "" }, { "first": "Leo", "middle": [], "last": "Lepp\u00e4nen", "suffix": "" }, { "first": "Elaine", "middle": [], "last": "Zosa", "suffix": "" }, { "first": "Matej", "middle": [], "last": "Ul\u010dar", "suffix": "" }, { "first": "Linda", "middle": [], "last": "Freienthal", "suffix": "" }, { "first": "Silver", "middle": [], "last": "Traat", "suffix": "" }, { "first": "Luis", "middle": [ "Adri\u00e1n" ], "last": "Cabrera-Diego", "suffix": "" }, { "first": "Matej", "middle": [], "last": "Martinc", "suffix": "" }, { "first": "Nada", "middle": [], "last": "Lavra\u010d", "suffix": "" }, { "first": "", "middle": [], "last": "Bla\u017e\u0161krlj", "suffix": "" }, { "first": "Andra\u017e", "middle": [], "last": "Mar-Tin\u017enidar\u0161i\u010d", "suffix": "" }, { "first": "Boshko", "middle": [], "last": "Pelicon", "suffix": "" }, { "first": "Vid", "middle": [], "last": "Koloski", "suffix": "" }, { "first": "Janez", "middle": [], "last": "Podpe\u010dan", "suffix": "" }, { "first": "Shane", "middle": [], "last": "Kranjc", "suffix": "" }, { "first": "", "middle": [], "last": "Sheehan", "suffix": "" } ], "year": null, "venue": "Proceedings of the EACL Hackashop on News Media Content Analysis and Automated Report Generation", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Senja Pollak, Marko Robnik\u0160ikonja, Matthew Purver, Michele Boggia, Ravi Shekhar, Marko Pranji\u0107, Salla Salmela, Ivar Krustok, Tarmo Paju, Carl-Gustav Linden, Leo Lepp\u00e4nen, Elaine Zosa, Matej Ul\u010dar, Linda Freienthal, Silver Traat, Luis Adri\u00e1n Cabrera- Diego, Matej Martinc, Nada Lavra\u010d, Bla\u017e\u0160krlj, Mar- tin\u017dnidar\u0161i\u010d, Andra\u017e Pelicon, Boshko Koloski, Vid Podpe\u010dan, Janez Kranjc, Shane Sheehan, Hannu Toivonen, Emanuela Boros, Jose Moreno, and An- toine Doucet. 2021. EMBEDDIA tools, datasets and challenges: Resources and hackathon contributions. In Proceedings of the EACL Hackashop on News Me- dia Content Analysis and Automated Report Gener- ation. Association for Computational Linguistics.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Language models are unsupervised multitask learners", "authors": [ { "first": "Alec", "middle": [], "last": "Radford", "suffix": "" }, { "first": "Jeffrey", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Rewon", "middle": [], "last": "Child", "suffix": "" }, { "first": "David", "middle": [], "last": "Luan", "suffix": "" }, { "first": "Dario", "middle": [], "last": "Amodei", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Techni- cal report, OpenAI.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Automatic keyword extraction from individual documents", "authors": [ { "first": "Stuart", "middle": [], "last": "Rose", "suffix": "" }, { "first": "Dave", "middle": [], "last": "Engel", "suffix": "" }, { "first": "Nick", "middle": [], "last": "Cramer", "suffix": "" }, { "first": "Wendy", "middle": [], "last": "Cowley", "suffix": "" } ], "year": 2010, "venue": "Text mining: applications and theory", "volume": "1", "issue": "", "pages": "1--20", "other_ids": {}, "num": null, "urls": [], "raw_text": "Stuart Rose, Dave Engel, Nick Cramer, and Wendy Cowley. 2010. Automatic keyword extraction from individual documents. Text mining: applications and theory, 1:1-20.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Keyphrase extraction from scholarly articles as sequence labeling using contextualized embeddings", "authors": [ { "first": "Dhruva", "middle": [], "last": "Sahrawat", "suffix": "" }, { "first": "Debanjan", "middle": [], "last": "Mahata", "suffix": "" }, { "first": "Mayank", "middle": [], "last": "Kulkarni", "suffix": "" }, { "first": "Haimin", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Rakesh", "middle": [], "last": "Gosangi", "suffix": "" }, { "first": "Amanda", "middle": [], "last": "Stent", "suffix": "" }, { "first": "Agniv", "middle": [], "last": "Sharma", "suffix": "" }, { "first": "Yaman", "middle": [], "last": "Kumar", "suffix": "" }, { "first": "", "middle": [], "last": "Rajiv Ratn", "suffix": "" }, { "first": "Roger", "middle": [], "last": "Shah", "suffix": "" }, { "first": "", "middle": [], "last": "Zimmermann", "suffix": "" } ], "year": 2020, "venue": "Proceedings of European Conference on Information Retrieval (ECIR 2020)", "volume": "", "issue": "", "pages": "328--335", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dhruva Sahrawat, Debanjan Mahata, Mayank Kulka- rni, Haimin Zhang, Rakesh Gosangi, Amanda Stent, Agniv Sharma, Yaman Kumar, Rajiv Ratn Shah, and Roger Zimmermann. 2020. Keyphrase extraction from scholarly articles as sequence labeling using contextualized embeddings. In Proceedings of Eu- ropean Conference on Information Retrieval (ECIR 2020), pages 328-335.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Introduction to modern information retrieval", "authors": [ { "first": "Gerard", "middle": [], "last": "Salton", "suffix": "" }, { "first": "J", "middle": [], "last": "Michael", "suffix": "" }, { "first": "", "middle": [], "last": "Mcgill", "suffix": "" } ], "year": 1986, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gerard Salton and Michael J McGill. 1986. Introduc- tion to modern information retrieval.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Rakun: Rank-based keyword extraction via unsupervised learning and meta vertex aggregation", "authors": [ { "first": "Andra\u017e", "middle": [], "last": "Bla\u017e\u0161krlj", "suffix": "" }, { "first": "Senja", "middle": [], "last": "Repar", "suffix": "" }, { "first": "", "middle": [], "last": "Pollak", "suffix": "" } ], "year": 2019, "venue": "International Conference on Statistical Language and Speech Processing", "volume": "", "issue": "", "pages": "311--323", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bla\u017e\u0160krlj, Andra\u017e Repar, and Senja Pollak. 2019. Rakun: Rank-based keyword extraction via unsuper- vised learning and meta vertex aggregation. In In- ternational Conference on Statistical Language and Speech Processing, pages 311-323. Springer.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Topical word importance for fast keyphrase extraction", "authors": [ { "first": "Lucas", "middle": [], "last": "Sterckx", "suffix": "" }, { "first": "Thomas", "middle": [], "last": "Demeester", "suffix": "" }, { "first": "Johannes", "middle": [], "last": "Deleu", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Develder", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 24th International Conference on World Wide Web", "volume": "", "issue": "", "pages": "121--122", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lucas Sterckx, Thomas Demeester, Johannes Deleu, and Chris Develder. 2015. Topical word importance for fast keyphrase extraction. In Proceedings of the 24th International Conference on World Wide Web, pages 121-122.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Joint keyphrase chunking and salience ranking with bert", "authors": [ { "first": "Si", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Chenyan", "middle": [], "last": "Xiong", "suffix": "" }, { "first": "Zhenghao", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Zhiyuan", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Jie", "middle": [], "last": "Bao", "suffix": "" } ], "year": 2020, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:2004.13639" ] }, "num": null, "urls": [], "raw_text": "Si Sun, Chenyan Xiong, Zhenghao Liu, Zhiyuan Liu, and Jie Bao. 2020. Joint keyphrase chunking and salience ranking with bert. arXiv preprint arXiv:2004.13639.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Attention is all you need", "authors": [ { "first": "Ashish", "middle": [], "last": "Vaswani", "suffix": "" }, { "first": "Noam", "middle": [], "last": "Shazeer", "suffix": "" }, { "first": "Niki", "middle": [], "last": "Parmar", "suffix": "" }, { "first": "Jakob", "middle": [], "last": "Uszkoreit", "suffix": "" }, { "first": "Llion", "middle": [], "last": "Jones", "suffix": "" }, { "first": "Aidan", "middle": [ "N" ], "last": "Gomez", "suffix": "" }, { "first": "Lukasz", "middle": [], "last": "Kaiser", "suffix": "" }, { "first": "Illia", "middle": [], "last": "Polosukhin", "suffix": "" } ], "year": 2017, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1706.03762" ] }, "num": null, "urls": [], "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Wikidata: A free collaborative knowledgebase", "authors": [ { "first": "Denny", "middle": [], "last": "Vrande\u010di\u0107", "suffix": "" }, { "first": "Markus", "middle": [], "last": "Kr\u00f6tzsch", "suffix": "" } ], "year": 2014, "venue": "Commun. ACM", "volume": "57", "issue": "10", "pages": "78--85", "other_ids": { "DOI": [ "10.1145/2629489" ] }, "num": null, "urls": [], "raw_text": "Denny Vrande\u010di\u0107 and Markus Kr\u00f6tzsch. 2014. Wiki- data: A free collaborative knowledgebase. Commun. ACM, 57(10):78-85.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Single document keyphrase extraction using neighborhood knowledge", "authors": [ { "first": "Xiaojun", "middle": [], "last": "Wan", "suffix": "" }, { "first": "Jianguo", "middle": [], "last": "Xiao", "suffix": "" } ], "year": 2008, "venue": "AAAI", "volume": "8", "issue": "", "pages": "855--860", "other_ids": {}, "num": null, "urls": [], "raw_text": "Xiaojun Wan and Jianguo Xiao. 2008. Single doc- ument keyphrase extraction using neighborhood knowledge. In AAAI, volume 8, pages 855-860.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Kea: Practical automated keyphrase extraction", "authors": [ { "first": "H", "middle": [], "last": "Ian", "suffix": "" }, { "first": "Gordon", "middle": [ "W" ], "last": "Witten", "suffix": "" }, { "first": "Eibe", "middle": [], "last": "Paynter", "suffix": "" }, { "first": "Carl", "middle": [], "last": "Frank", "suffix": "" }, { "first": "Craig G Nevill-Manning", "middle": [], "last": "Gutwin", "suffix": "" } ], "year": 2005, "venue": "Design and Usability of Digital Libraries: Case Studies in the Asia Pacific", "volume": "", "issue": "", "pages": "129--152", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ian H Witten, Gordon W Paynter, Eibe Frank, Carl Gutwin, and Craig G Nevill-Manning. 2005. Kea: Practical automated keyphrase extraction. In Design and Usability of Digital Libraries: Case Studies in the Asia Pacific, pages 129-152. IGI global.", "links": null } }, "ref_entries": { "FIGREF0": { "text": "Preprocessing pipeline used for the document normalization and cleaning.", "type_str": "figure", "uris": null, "num": null }, "FIGREF1": { "text": "& BERT + BiLSTM-CRF & TF-IDF(tm) 0.2648 0.5681 0.3612 0.1699 0.7040 0& BERT + BiLSTM-CRF & TF-IDF(tm) 0.3588 0.6206 0.4547 0.2107 0.6912 0.3230 Russian TF-IDF 0.1764 0.2314 0.2002 0.1663 0.3350 0.2223 TNT-KID 0.7108 0.6007 0.6512 0.7038 0.6250 0.6621 BERT + BiLSTM-CRF 0.6901 0.5467 0.5467 0.6849 0.5643 0.6187 TNT-KID & TF-IDF(tm) 0.4519 0.6293 0.5261 0.2981 0.6946 0.4172 BERT + BiLSTM-CRF & TF-IDF(tm) 0.4157 0.5728 0.4818 0.2753 0.6378 0.3846 TNT-KID & BERT + BiLSTM-CRF 0.6226 0.6375 0.6300 0.5877 0.6707 0.6265 TNT-KID & BERT + BiLSTM-CRF & TF-IDF(tm) 0.4622 0.6527 0.5412 0.2965 0.7213 0& BERT + BiLSTM-CRF & TF-IDF(tm) 0.3318 0.7852 0.4666 0.2124 0.8672 0.3414", "type_str": "figure", "uris": null, "num": null }, "TABREF0": { "type_str": "table", "num": null, "content": "
Dataset Unique tags Type of tags
Croatian21,165Constructed
Estonian52,068Provided
Russian5,899Provided
Latvian4,015Constructed
", "html": null, "text": "." }, "TABREF1": { "type_str": "table", "num": null, "content": "", "html": null, "text": "Distribution of tags provided per language. The media houses provided tagsets for Estonian and Russian, while the tags for Latvian and Croatian were extracted from the train set." }, "TABREF2": { "type_str": "table", "num": null, "content": "
Avg. TrainAvg. Test
Dataset Total docs Croatian 35,805126,68432,223438.50 3.540.321.193582464.39 3.530.341.26
Estonian18,49759,24210,750395.24 3.810.652.777,747411.59 4.090.693.12
Russian25,3065,95313,831392.82 5.660.764.4411,475335.93 5.430.794.33
Latvian24,7744,03613,133378.03 3.230.531.6911,641460.15 3.190.551.71
", "html": null, "text": "Total kw. Total docs Doc len Kw. % present kw. present kw. Total docs Doc len Kw. % present kw. Present kw." }, "TABREF3": { "type_str": "table", "num": null, "content": "", "html": null, "text": "Media partners' datasets used for empirical evaluation of keyword extraction algorithms." }, "TABREF4": { "type_str": "table", "num": null, "content": "
", "html": null, "text": "Results on the EMBEDDIA media partner datasets." } } } }